HPC-Driven Oceanographic Predictions with Graph Neural Networks (GNNs) and Gated Recurrent Units (GRUs)

EasyChair Preprint 14943

Abstract

In this work, we utilized the high-performance computing (HPC) capabilities of the Vienna Scientific Cluster (VSC5) to develop and validate advanced AI models for oceanographic forecasting, with a focus on predicting Significant Wave Height (SWH). Using the computational power of VSC5, particularly its NVIDIA A100 GPUs, allowed us to process and analyze over 500 million data points. The most promising model, a Graph Neural Networks - Gated Recurrent Unit GNN-GRU hybrid, was trained to generate six hourly forecasts for SWH and achieved a Mean Absolute Error (MAE) of 0.0071 and an R-squared (R2) value of 0.98 against test data, demonstrating high accuracy and efficiency. The first validation results outside the training period are promising and efforts to refine the model are on-going. This work highlights the importance of HPC to the advancement of oceanographic forecasting, by enabling the processing of extremely large datasets and the creation of models that are up to the demanding standards of the marine sector. Subsequent efforts will center around enhancing these models, mitigating detected biases, and creating an operational deployment framework that can facilitate prompt decision-making in maritime operations.

Keyphrases: CDS, CMEMS, EuroCC, GNN-GRU, HPC, significant wave height

@booklet{EasyChair:14943,
  author    = {Paraskevi Vourlioti and Theano Mamouka and Maria Banti and Charalampos Paraskevas and Stylianos Kotsopoulos and Vasileios Alexandridis and Georgia Kalantzi},
  title     = {HPC-Driven Oceanographic Predictions with Graph Neural Networks (GNNs) and Gated Recurrent Units (GRUs)},
  howpublished = {EasyChair Preprint 14943},
  year      = {EasyChair, 2024}}